Visual review
Step
- Does
- looks at frames across each episode and records a grade and a note per scene the metrics may have misjudged.
- Passes when
- every rendered scene's contact sheet has been looked at and
visual_review.jsoncovers every scene you have an opinion on.
Inputs. A collected run dir from render-and-grade. Tools. ffmpeg, collect.py.
- Open every
media/SCENE_sheet.jpg(12 frames across the episode). Look for jitter, explosions, interpenetration, launched or floating bodies, and bodies vanishing into geometry. - When a sheet looks off, pull single frames:
ffmpeg -ss 7.5 -i "$RUN/out/SCENE/clip_raw.mp4" -frames:v 1 frame.jpg - Before calling a body lost, find it in the poses: a fixed camera cannot see into a bin behind a deck, and the example’s own check may count it. Compare doubtful metrics with another solver on the same scene.
- Write
$RUN/visual_review.json:
then run{"crane_wrecking_ball": {"grade": "ok", "note": "the ball knocks the building down as scripted; nothing launches or sinks"}}collect.pyandbuild_gallery.pyagain.
Rules. A visual note can raise a grade; it never hides an automatic unstable or failed. Say what you saw, at
what time, in which part of the frame.
Outputs. visual_review.json and a regenerated results.json and gallery.
Source: agents/steps/visual-review/README.md, the copy agents read.
Steel docs and tutorials are licensed under Apache-2.0. Third-party notices. "Skild", "Skild AI" and the Skild logo are trademarks of Skild AI and are not licensed under Apache-2.0.